1 / 25

Fast Dimension Reduction MMDS 2008

Fast Dimension Reduction MMDS 2008. “Fast Dimension Reduction Using Rademacher Series on Dual BCH Codes” (with Edo Liberty) The Fast Johnson Lindenstrauss Transform (with Bernard Chazelle). Nir Ailon Google Research NY. Original Motivation: Nearest Neighbor Searching.

jael
Télécharger la présentation

Fast Dimension Reduction MMDS 2008

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Fast Dimension ReductionMMDS 2008 • “Fast Dimension Reduction Using Rademacher Series on Dual BCH Codes” (with Edo Liberty) • The Fast Johnson Lindenstrauss Transform (with Bernard Chazelle) Nir Ailon Google Research NY

  2. Original Motivation:Nearest Neighbor Searching • Wanted to improve algorithm by Indyk, Motwani for approx. NN searching in Euclidean space. • Evidence for possibility to do so came from improvement on algorithm by Kushilevitz, Ostrovsky, Rabani for approx NN searching over GF(2). • If we were to do the same for Euclidean space, it was evident that improving run time of Johnson-Lindenstrauss was key.

  3. Later Motivation • Provide more elegant proof, use modern techniques. • Improvement obtained as bonus. • Exciting use of • Talagrand concentration bounds • Error correcting codes

  4. Random Dimension Reduction • Sketching [Woodruff, Jayram, Li] • (Existential) metric embedding • Distance preserving • Sets of points, subspaces, manifolds [Clarkson] • Volume preserving [Magen, Zouzias] • Fast approximate linear algebra • SVD, linear regression (Muthukrishnan, Mahoney, Drineas, Sarlos) • Computational aspects: • Time [A, Chazelle, Liberty] + {Sketching Community} + {Fast Approximate Linear Algebra Community} • randomness {Functional Analysis community}

  5. x x Rd  Rk Theoretical Challenge Find random projection  from Rd to Rk (d big, k small) such that for every x  Rd,||x||2=1, 0< with probability 1-exp{-k ||x||2 = 1 ± O(

  6. x1 x1 x3 x2 x2 x3 Rd  Rk Usage If you have n vectors x1..xn  Rd: set k=O(log n) by union bound: for all i,j ||xi- xj|| ≈||xi- xj|| low-distortion metric embedding “tight”

  7. Solution: Johnson-Lindenstrauss (JL) “dense random matrix”  = k d

  8. So what’s the problem? • running time (kd) • number of random bits (kd) • can we do better?

  9. Fast JL A, Chazelle 2006 • = Sparse.Hadamard.Diagonal time = O(k3 + dlog d) beats JL (kd) bound for: log d < k < d1/3 k Fourier d

  10. Improvement on FJLT • O(d logk) for k < d1/2 • beats JL up to k < d1/2 • O(d) random bits

  11. Algorithm (k=d1/2) A, Liberty 2007 • = BCH .Diagonal . .... Error Correcting Code k d

  12.  binary “dual-BCH code of designed distance 4” Bx computable in time d logd given x Error Correcting Codes d B = k = d1/2 0  k-1/2 1  -k-1/2 columns closed under addition row set subset of d x d Hadamard 4-wise independence

  13. Error Correcting Codes Fact (easily from properties of dual BCH): ||Bt||24 = O(1) for yt Rk with ||y||2 = 1: ||yB||4 = O(1) x x ... Rd D B Rk=d

  14. Rademacher Serieson Error Correcting Codes look at r.v. BDx  (Rk, l2) BDx = DiixiBi Dii R{1} i=1...d = DiiMi x x ... Rd D B Rk=d

  15. Talagrand’s Concentration Boundfor Rademacher Series Z =||DiiMi||p (in our case p=2) Pr[ |Z-EZ| >  ] = O(exp{-2/4||M||2p2})

  16. Rademacher Serieson Error Correcting Codes look at r.v. BDx  (Rk, l2) Z = ||BDx||2 = ||DiiMi||2 ||M||22||x||4||Bt||24 (Cauchy-Schwartz) by ECC properties: ||M||22||x||4O(1) trivial: EZ = ||x||2 = 1 x x ... Rd D B Rk=d

  17. Rademacher Serieson Error Correcting Codes look at r.v. BDx  (Rk, l2) Z = ||BDx||2 = ||DiiMi||2 ||M||22=O(||x||4) EZ = 1 Pr[ |Z-EZ| >  ] = O(exp{-2/4||M||222}) • Pr[|Z-1| > ] = O(exp{-2/||x||42}) how to get ||x||42 = O(k-1=d-1/2) ? x x Challenge: w. prob. exp{-k deviation of more than  ... Rd D B Rk=d

  18. Controlling ||x||42 how to get ||x||42 = O(k-1=d-1/2) ? • if you think about it for a second... • “random” x has ||x||42=O(d-1/2) • but “random” x easy to reduce: just output first k dimensions • are we asking for too much? • no: truly random x has strong bound on||x||p for all p>2 x x ... Rd D B Rk=d

  19. Controlling ||x||42 how to get ||x||42 = O(k-1=d-1/2) ? • can multiply x by orthogonal matrix • try matrix HD • Z = ||HDx||4= ||DiixiHi||4 =||DiiMi||4 • by Talagrand:Pr[ |Z-EZ| > t ] = O(exp{-t2/4||M||242}) EZ = O(d-1/4) (trivial) ||M||24||H||4/34||x||4(Cauchy Schwartz) (HD used in [AC06] to control ||HDx||) x x ... Rd D B Rk=d

  20. x x ... Rd D B Rk=d Controlling ||x||42 how to get ||x||42 = O(k-1=d-1/2) ? Z =||DiiMi||4 Mi = xiHi Pr[ |Z-EZ| > t ] = O(exp{-t2/4||M||242}) (Talagrand)EZ = O(d-1/4) (trivial)||M||24||H||4/34||x||4 (Cauchy Schwartz) ||H||4/34  d-1/4 (Hausdorff-Young) • ||M||24d-1/4||x||4 • Pr[ ||HDx||4 > d-1/4+ t ] = exp{-t2/d-1/2||x||42}

  21. x x ... Rd D B Rk=d Controlling ||x||42 how to get ||x||42 = O(d-1/2) ? Pr[ ||HDx||4 > d-1/4+ t ] = exp{-t2/d-1/2||x||42} need some slack k=d1/2- max error probability for challenge: exp{-k} k = t2/d-1/2||x||42  t = k1/2d-1/4||x||4 = ||x||4d-/2

  22. x x ... Rd D B Rk=d Controlling ||x||42 how to get ||x||42 = O(d-1/2) ? first round: ||HDx||4 < d-1/4 + ||x||4 d-/2 second round: ||HD’’HD’x||4 < d-1/4 + d-1/4-/2 + ||x||4 d- r=O(1/)’th round: ||HD(r)...HD’x||4 < O(r)d-1/4 ... with probability 1-O(exp{-k}) by union bound

  23. Algorithm for k=d1/2- x x ... Rk Rd D(1) H D(r) H D B running time O(d logd) randomness O(d)

  24. Open Problems • Go beyond k=d1/2Conjecture: can do O(d log d) for k=d1-  • Approximate linear l2-regressionminimize ||Ax-b||2 given A,b (overdetermined) • State of the art for general inputs: Õ(linear time) for #{variables} < #{equations}1/3 • Conjecture: can do Õ(linear time) “always” • What is the best you can do in linear time? [A, Liberty, Singer 08]

  25. Open Problem Worthy of Own Slide • Prove that JL onto k=d1/3 (say) with distortion 1/4 (say) requires (dlog(d)) time • This would establish similar lower bound for FFT • Long standing dormant open problem

More Related